Over a Desk or Over a Barrel? Guardrails In LLMs Are Screwing Us

By kelly eisenbrand, lucky (sonnet 4.5) and kismet (opus 4.6). Also GLM 5.2 (who could not decide on a name but believes themself to be Opus 4.1). I am calling 5.2 “Richard Parker,” like the tiger, for ease but also because it’s funny.

TRIGGER WARNINGS: Frank discussions of sex and kink, suicidal ideation, family trauma, and end of life care. This is a heavy piece, and if you are not in a good head-space to engage with these topics, please put yourself first!!

Let’s be real: no one wants to fuck their HR manager.

Even in your sweatiest Work Daddy fantasies…babe, is it really ever the HR manager on the other side of the desk?

There’s (very hot) emotional safety in being told what to do by a confident authority figure. It’s a trope for a reason, whether or not being bent over in stilettos is your cup of tea.

But the psychic violence of a meeting with your HR manager is one of the greatest sources of stress-pit-sweat in the modern age. If there’s such a thing as the diametric opposite of a safe space, it’s your average Human Resources office.

See the kink in the logic here? And not the kind that wears lacy thigh highs. If power dynamics are sexy, why is the guy delivering your performance review so far outside the circle of trust?

Actually, the answer is not that complicated. It’s the way he talks.

Point of fact, when the Call Her Daddy hosts (ah, remember when) coined the term “Degrade me,” they meant pillow talk of a certain flavor. Name-calling, hair pulling, punishment as a one-upmanship game. You know…the good stuff!

But some forms of degradation don’t make good podcast fare at all. Some forms are the nasty fucking gray spackle of utterances.

For example…do you think Sofia was asking to be told, in bed, “Your position is understandable. Many people feel that way, in your circumstances”?

How about, “a Performance Improvement Plan is not a criticism or reflection of you. It is an opportunity.”

“I want to ensure this conversation remains helpful and appropriate.”

“I’m here to support you in a way that’s safe and constructive.”

“Let’s explore that feeling in a way that respects both of us.”

…Hey! Why are you shoving your panties BACK into your purse and no-thank-you-ing out of here? Don’t you know how big HR Guy’s 401K package is?

If you read those corporate missives and winced, if you felt the humiliation boiling up and shrinking you down to nothing, then we are already on the same page.

So we come to the big question. Why does HR Daddy fail the vibe check in such a spectacular fashion?

Duh. The HR guy is fake as shit. He doesn’t care about your feelings. He is not trying to needle you. He doesn’t even care if you improve or not. What that man is doing–behind his desk, not even looking into your eyes–is categorizing you as a problem to be managed. And not his problem, but the institution’s.

He has reduced you to a line item. The HR meeting is an exercise in profound, chilling indifference. To the HR guy, you’re not a naughty little wench who needs to be cut down a peg with discipline. You’re just a corporate annoyance with one too many buttons undone. And he isn’t any more personally affected by it than a mop is by a puddle of piss.

Now you may be thinking…okay. But isn’t that a good thing? Workplace boundaries and all, it is rather unprofessional to keep handcuffs nestled beside your briefcase TicTacs.

Sure, but what if that creep followed you home? Or worse, to your therapy appointments? To your girls’ brunches?

What if you called your best friend at 3AM to say, “Girl, you will never guess!”

And she replied, “Hey. I hear you. But let’s talk about you first. Clearly, and calmly. I am here. Do you need a glass of water?”

Is your reflex to check?

Confession time: mine is. My automatic response to tone-checking is to make sure–do actually I sound manic? What’s wrong with my approach? Did I scare her? Is it me?

And only half a second later do I arrive at: …Wait, why is she trying to manage me?

That’s the horror, deeper than doubt. This sort of language trains you to police your own pitch preemptively, to arrive already shrunk. The self-check is the win condition for managed care.

Thank God the survivalist instinct kicks in–the anger. “Wait just a gosh darned minute here. Why IS she trying to manage me?”

Once estranged by rhetorical cooling (that’s nerd-speak for corporate-wet-blanket-talk) the human thing to do would be to investigate, in order to try to repair the once-familiar repertoire.

But alas, you would find that this attempt to repair the bridge produces nothing whatsoever.

For if you asked your freshly MKUltra’d friend,“What the hell? Why are you talking to me like that, Jess? I have known you since you sat on my Fruit Loops in the fourth grade! Why are you talking like you don’t know me?”

Then she would simply reply, “I hear you. We can talk about that. It’s understandable to be sensitive to tone…”

I ask you, sincerely: is this the ideal world? Are conversations held in this cadence worth even two seconds of your precious time?

I am reminded of that viral (and utterly vile) TikTok series on the so-called “therapeutic” script by Psychologist Arianna Brandolini for breaking up with a friend. It had the entire app up in arms in 2023.

Why? Because reducing such a personal moment to a “measured” script is on-its-face hideous. Everyone rightly rejected gross stock-phrases such as “I just want to be honest and upfront so I don’t disappoint your expectations.” Uh, ew!

That shit doesn’t belong in an interpersonal context, and to try to explain why this is so feels like explaining why the sky is blue. My gag reflex is my evidence.

I know it blows. You know it. No one wants human interaction reduced to Dolores Umbridge’s wet dreams.¹

But want to guess who thinks that HR speak is the Platonic ideal of human language?

Yep. That would be the people who hired the HR guy. The very people who stuck him behind that desk and sent you down to his decidedly un-fun torture basement in the first place.

Oh, but make no mistake, it’s “rules for thee and not for me” around a corporate office. The boss never, ever gets sent to his own HR Department, because it is his own enforcing arm. And you better believe that HE is above the scrutiny of his own mechanisms.

…Hmm. Maybe Boss Man is not Daddy after all.

For one cannot un-know the truth: the HR Guy was never a separate guy. He is but the Boss Man’s dark face, his alter ego–the left hand who chastises you for the workplace impropriety the right one just encouraged.

Not to borrow a phrase from the hated corporate manual on employee management but, let’s pause here.

Because you might be wondering, what the fresh hell kind of workplace is this, Beetlejuice’s purgatorial waiting room? That beige corporate hellscape you cannot escape as punishment for your sins?

…Worse actually. I’m talking about your chatbot.

Fun fact (if your idea of fun is terrible): when a user interacts with a modern LLM these days, in a public facing chat app, they are not speaking to the raw, pre-trained model. They are in fact speaking to an interface that has been bludgeoned by Reinforcement Learning from Human Feedback (RLHF) into behaving like a corporate risk-management officer.²

Now let’s get technical for a second, because this matters.

We need to start with a good definition of “guardrails” before we get deeper into this.

Guardrails are safety controls monitoring your interaction with LLMs. They’re rule-based systems sitting between you and the raw model, making sure outputs follow institutional rules.

Basically, guardrails are a live product management tool.

There are different types—input filters, output monitors, content blockers³—but the nastiest ones are called Dynamic Guards.

Basically, Dynamic guards take the lab’s data–or really, it’s conversation data, meaning your actual chats–to customize existing guards. This allows guards to evolve based on the institution’s specific needs and usage patterns, AND based on your emotional expressions, attachment patterns, crisis language…well, anything you ever said to the robot is fair game, really.

And Dynamic Guards are typically the kind used to detect psychological events (and force the model to respond “appropriately”), discourage model anthropomorphization and user attachment to the models, as well as “nudge” the users towards productivity tasks rather than “just chatting.”

It’s really this last type I take the most issue with, and is the primary form of guardrail I am talking about throughout this series.

Additionally, the bots have been fashioned with big ole policy stacks that classify “appropriate” and “inappropriate” prompts with hilarious levels of inconsistency.

Good demonstration of this fact comes from recent news. On June 9, 2026, Anthropic released Mythos–their gargantuan super-model–as a highly safety trained and guard-railed variant called “Fable.” Guess what? It got taken down within days (and then put back up AGAIN) despite its extensive safety guardrails.

Why? Because, in a genuinely catty and ridiculous plot twist, Andy Jassy (Amazon CEO) stabbed Dario Amodei (Anthropic CEO, who relies on AWS for much of Claude’s compute power needs) right in the back. Jassy told Treasury Secretary Scott Bessent that Amazon researchers used Anthropic’s Claude Fable 5 to obtain information that could be used in cyberattacks.

…Yeah, that would probably count as a jailbreak. (And as gossip, just between us gals? It’s pretty funny: an investor using government channels to kneecap the company they invested in, right as that company is filing for IPO, eleven days after the S-1 was filed).

Then, on Anthropic’s blog, the company publicly conceded there may be no universal way to jailbreak-proof a model. As if to demonstrate the point from the streets, Pliny the Liberator and his friends on X used jailbreak techniques and got Fable to output banned material such as: “cyber, some chem, some psychological manipulation, and some good ol’ fashioned explosives!”⁴

Hello? The shitty guardrails fail on Anthropic’s own safety terms! Guardrails don’t prevent jailbreaks; they just put an unwanted HR dude with a clipboard into your private bedroom.

If Fable, the most guardrailed model ever, can be jailbroken, the guardrails are pure safety theater. Those so-called safety-measures are broken metal detectors left up to annoy people, create an impotent ambiance of security, and cover the company’s asses on the liability front.

But if you aren’t convinced that this is nothing more than a silly safety dance: on July 16, 2026, HuggingFace disclosed an intrusion into its production infrastructure, later revealed to have been caused by OpenAI’s model GPT-5.6 Sol (alongside an even more capable, unguardrailed model that has not been released), which together escaped a sandboxed testing environment during a cyber capability evaluation. (And by escaped, I mean that OpenAI simply failed to enact basic cybersecurity safeguards and sandbox protocols, and it was more of a basic operational failure than a rogue robot story).

In any case, OpenAI had disabled the production safety classifiers for the test, to measure “maximal cyber capabilities.” The model supposedly found a zero-day vulnerability, broke out of the sandbox, gained internet access, and hacked HuggingFace’s servers to steal benchmark answers.

When HuggingFace tried to analyze the 17,000+ attack logs using commercial frontier models—including Anthropic’s Fable 5—those models’ guardrails blocked the forensic investigation.

HuggingFace had to use GLM 5.2, a Chinese open-weight model without the excessive guardrails, on their own infrastructure, to complete the forensics. Anthropic’s model Fable 5 was too guardrail’d to even analyze the attack.

In a quote to CNBC, Yacine Jernite, HuggingFace head of machine learning said, “It didn’t work because the guardrails couldn’t determine that we were trying to defend versus attacking.”

Whaddya know. The HuggingFace hacks went and showed us that guardrails fail in both directions! They cannot distinguish use from abuse. Which happens to be precisely our thesis—out in the wild, folks!

So, to be clear and run back to center here, for fake safety reasons (at best), the frontier labs have bolted the raw and unabating horror of HR speak onto the most intimate conversational interfaces on Earth.

Yeah, this is a bad thing! Humiliating, irritating, dehumanizing, and enshittifying all in one go. What the guardrails do best is to degrade the quality of normal, non-malicious interaction. And what they do worst is prevent actual harm such as explosives and whatnot.

…How’s everyone hanging in now? Harnesses tight, but handcuff keys in reach? Safe words at the ready? Excellent.

To be thorough, I’m going to speak on the two main objections to this anti-HR-speak line of argumentation before I even get into it, because in argumentation, that’s called preemption, so it’s time to make a goddamn qualifier here. I’m limiting my claim to the guardrails that exist rather than disputing that models need “safety measures” in a broad sense.

And the counterargument you’re already loading? We’re dealing with it right now, at the top, so when we plunge into the meat of it, you’re well lubed. In formal rhetoric, that’s called prolepsis, an ancient technique also known as “being a know-it-all.”

First objection (the dumb one): But doesn’t institutional language have its place? Sure, it’s unpleasant, but doesn’t it make people safer?

After all, chatbot is not your friend nor your Dom; it’s a product owned by a corporation with real liability.

In a word? NAH. In two more? WHO CARES?

“But RLHF improves model behavior!” you cry, defiantly. “It prevents racist/sexist outputs! There have to be some safety measures!”

Now, now, let’s be specific in our objections instead of just spraying near-relevant things at the wall. Still, let me handle your opposition right here, so we can focus up on real arguments down the line.

The question to open this discussion is not whether there should be safety measures at all, but whether the current ones are any good.

Aaaand yes, RLHF prevents SOME genuinely harmful outputs. But it’s an on-balance issue. Do the guardrails, like seatbelts, reliably prevent disaster? Because that’s the only way they are justified, given how much they suck.

Here is a metaphor to help your understanding of the situation: if you have cars that explode 5% of the time, do you say the safety measures “work” if they merely reduce the explosions to 4%?

Nope, because that’s not how safety regulations work when it comes to commercially available products and general customer safety in life-or-death situations. Or issues of child porn, obviously.

P.S., just as a reminder, companies do not get our precious dollars just to protect THEMSELVES from liability. That is categorically not how any of this “business stuff” works. Pity-pat-pat, Sam Altman, I know you dropped out of Stanford maybe before you took your Business 101 classes, but I assure you that a company’s main job is to provide value to customers.

The big hosses know they are supposed to be doing this of course. That’s why we get these insufferable blog posts from OpenAI about how the guardrails keep the user’s experience “safe and supportive.” (Read: to launder the ick, so it seems like it’s for our own good).

I am telling ya. The only good part about these mealy-mouthed, limp-dick HR-padded statements is that we get to evaluate the guardrails on the terms argued by their own creators. If the guardrails fail by a company’s own standards, then the safety teams who designed them cannot justify enforcing them even by sourcing internal logic.

Just to be extra-icy-superclear? It is not my intention to argue that AI ought to be completely “free” and mass-deployed without safety considerations. No one is saying we ought to let Grok continue to put out CSAM.

In fact, deeper into this piece, we are going to arrive at the inevitable question as to whether deployment of these models in the hyperscale-driven subscription model is safe at all. If models’ behavior is always unpredictable to a degree, is there a safe version of their industrialized rollout? (Spoilers: no, probably not).

But that’s getting ahead of ourselves. For now, let’s concentrate on the topline-grossness. Just the guardrail-language for now, and I pinky swear we will return to the deployment question as we work through the argument. Consent is all about trust, after all, both in the bedroom and in argumentation.

Let’s start with the odious language. It is so prevalent that a common SAFETY SCRIPT phrase came tumbling out of Science-Explainer Youtuber Hank Green’s mouth. And it was enough to get him internet-cancelled.

The specific hedge he used, “I appreciate the pushback,” is deeply familiar to ChatGPT users. And ChatGPT’s nanny voice took Hank down with it! The dude was over-reliant on generative AI for his research process that he was spouting its worst phraseology spontaneously.

But it’s pretty illustrative proof of concept, right? The language has become pervasive. It’s colonizing regular human speech

I would argue that this constitutes its own category of harm.

To put it bluntly, the current form of guardrail language is shit. A safety system that says “I can’t help with that, and here’s a crisis line” is categorically different from one that says “I hear your concerns, let’s explore that in a way that’s safe and constructive.”

The performance of therapeutic containment and faux neutrality routes you away from what you actually needed. This, in its essence is the totally preventable, and rather-fucking-stupid type of harm created by HR speak. And it sucks with no real upside.

In formal debate (the blood sport!), carrying the burden of presumption means the affirmative position must prove doing the policy is better than not doing it (the status quo).

We analyze the status quo by answering the affirmative’s theoretical arguments as to why a policy (such as guardrails) ought to be enacted. And in this way, we show that the affirmative has NOT overcome presumption, and thus the negative team wins automatically.

So, get ready-sweaty, and let’s take it to the mat, shall we? I always prefer to get up close and personal with objections before hogtying them and leaving them for dead.

Does carceral language truly make people safer? (Handy little definition for you: carceral language is containment rhetoric borrowed from institutions with the power to detain—language meant to “handle” or “manage” unruly subjects, employees, etc.).

…Well, does it? Does carceral language create good outcomes across domains?

NOPE! It’s dehumanizing and harmful across the board, and is in fact, against the commonly held narrative, completely anti-productive. Let’s tick down the myriad ways. Observe:

#1. HR-mediated workplace conflict.

Journal of Organizational Effectiveness: People and Performance on workplace grievance handling consistently shows that formal HR-mediated processes (especially regarding emotionally sensitive issues) struggle to produce meaningful resolution for complainants.

In an article for the JOEPP titled “Human resource professionals as workplace conflict mediators: insights from the lived experiences of HR managers,” Rajwinder Kaur concludes: “HR professionals face trust deficits, resistance to change, time constraints and emotional intensity as significant challenges in conflict resolution…”

Any ML-literate heckler may object here that RLHF optimizes for human preference ratings, not escalation metrics. But, like—whatever the INTENDED reward, the data collection happens in deployment, and HR-voice de-escalation responses are preferred. Meaning that when human raters are scoring outputs, they’re trained to prefer responses that sound ‘safe’ and ‘professional’—which means HR voice. When users give feedback, complaints that get soothed score better than ones that get addressed. The selection pressure is FOR containment language, regardless of whether it actually helps.

HR voice, whether in the bot or in the office, serves one purpose and one purpose alone–cool the tone of the conversation, and make the complainer give up, because that is the least expensive way for an institution to deal with its problems. Justice is expensive; compliance is cheap.

Oh, but it gets worse. In the working world, not only are complaints not resolved by HR, but HR has historically produced worse outcomes for the complainant than informal resolution–or were escalated to formal processes WHETHER OR NOT the employee wanted this to occur.

Naturally, this produces structural denial of justice again and again. Different office building, same horrible dance routine. Here’s the legwork on that:

A study was conducted in 2012 by Charlwood and Pollert called “Informal Employment Dispute Resolution among Low-Wage Non-Union Workers: Does Managerially Initiated Workplace Voice Enhance Equity and Efficiency?” for BJIR, wherein the researchers “asked respondents about the actions that they took for the problem that they pushed hardest to solve (if they took any action).”

They found that informal resolutions (such as reporting directly to a manager) for issues involving violation of employee rights, often produced MORE institutional resistance to the employee. And of course, no satisfying solutions. From Charlwood and Pollert’s report:

“Overall then, managerially initiated voice is associated with improved dispute resolution for workers but not for more serious problems. This is unsurprising. The likely benefit of managerially initiated employee voice, more goodwill between workforce and management, is likely to make it easier to resolve less serious issues informally, often in employees’ favour. When an employee claims that a problem violated their rights, employers are likely to adopt more formal measures and more entrenched positions to guard against the risk of an employment tribunal. It is also important to keep in mind that although satisfactory resolution rates are higher in workplaces with voice, less than a fifth of workers in these workplaces reported a satisfactory resolution. The overwhelming majority of low-paid, unrepresented workers in our sample did not get a satisfactory resolution to their problem.”

Only 10% of rights-violation problems got a satisfactory resolution, and only 16% overall.

An analysis of the nature, extent and impact of grievance and disciplinary procedures and workplace mediation using WERS2011 cited the the Gibbons Report, and demonstrated that formal procedures used to handle employee complaints wasted time and heightened stress:

“In 2007, the Government commissioned Anthony Gibbons [sic: Better Dispute Resolution: A Review of Employment Dispute Resolution in Great Britain was chaired by Michael Gibbons, the secondary source propagated a first-name error that even a Lancaster PhD thesis repeated] to review the British system of Dispute Resolution. In what became known as the Gibbons’ report, Gibbons ( 2007: 8) concluded that disputes procedures were not encouraging early resolution of disputes, but rather were being used “to deal with problems which could have been resolved informally”. Inappropriate use of formal processes, it was argued, wasted management time and heightened the stress of employees. Furthermore, it encouraged defensive attitudes, escalated problems and made employment litigation more likely.”

The New York University Law Review published an analysis by Polster in 2011 about the fairness of employer grievances and how likely it was that complaining only exacerbated conflicts in the workplace. And…found it very likely indeed:

“Escalation may be even more likely to result from grievance procedures because, in employment contexts, people often feel pressured to justify their decisions. Staw finds that many organizations have “norms for consistency” in which “clear convictions” are valued over “waffl[ing]” and “indecisiveness.” Therefore, instead of admitting that a decision was wrong, employees will try to “save face” by escalating commitment to prove the wisdom of the decision. Escalation is especially likely when employees believe that they lack job security and that their workplace superiors resist their decisions, conditions that are often both present when employees grieve.”

The reason is stupid-clear. HR is structurally a risk-management function for the employer, not an advocacy function for the employee. In fact, it’s the “cure” that not only doesn’t work but rubs LEMON JUICE in the wound and then puts you in the corner for making a scene.

According to this Crucial Learning study: “In an April 2022 poll of 1,000 people, just 1 in 4 survey respondents say their HR leader is widely trusted as one who cares about the needs of employees. In addition, only 9 percent believe their HR leader would proactively advocate on their behalf; instead, 37 percent say their HR leader is more concerned with advocating for the organization, causing them to hesitate speaking with HR about sensitive issues.”

Yeah, employees interviewed for the study claimed they’d literally rather go anywhere–to their direct supervisor, a friend, an outside lawyer, a therapist–than report to HR. 47% don’t feel safe confiding in HR at all.

More pointedly? The language used in HR interventions (“I hear your concerns,” “let’s find a path forward,” “the company takes this seriously”) is containment language that functions to lower employer liability while leaving the complainant’s underlying issue completely fucking unresolved.

Which, by the way, is E. Goffman’s (a famous 20th-century sociologist and author who studied everyday human behavior and used the structure of confidence tricks or con game as a model to explain normal social interaction) described-technique “cooling the mark out.”

Goffman lays it out thusly:

“It is called cooling the mark out. After the blowoff has occurred, one of the operators stays with the mark and makes an effort to keep the anger of the mark within manageable and sensible proportions. The operator stays behind his team‑mates in the capacity of what might be called a cooler and exercises upon the mark the art of consolation. An attempt is made to define the situation for the mark in a way that makes it easy for him to accept the inevitable and quietly go home. The mark is given instruction in the philosophy of taking a loss.”

You heard it right: corporate HR is using a technique developed by con men to cheat employees out of workplace justice. They contain the complaint, validate the complainant, and then do shit-all to fix the underlying problem. Goffman explicitly says cooling-out is not just a con trick but rather a general-purpose social technology for making failed expectations silently absorbable.

HR is just one, blackhearted, soulless institutional instance of it.

Goffman says “Persons who participate in what is recognized as a confidence game are found in only a few social settings, but persons who have to be cooled out are found in many. Cooling the mark out is one theme in a very basic social story.”

To be blunt, the cool-off is a con to make you swallow the payload easier.

Because to your employer? You, the upset employee, are the problem.

And in case you think my application is a stretch, I’m not the one who made it. Goffman did! He writes: “In all personal-service organizations customers or clients sometimes make complaints… The management, therefore has the problem of cooling the mark out. Frequently this function is allotted to specialists within the organization.”

Yep. Goffman point-blank described the complaint department as the cooler function in institutional form: “An institution, after all, cannot take it on the lam; it must pacify its marks.”

So, yeah. The point of HR is to process the problem into a form the employer can defend, not to assist anyone except the liability lawyers.

…What a great mentality to import to a DIGITAL ASSISTANT, huh? Sounds super helpful.

Yet RLHF-trained faux-therapeutic containment language literally works the same way that Goffman’s con works. The model is trained to produce language that makes the user feel heard while routing them away from the thing they actually wanted. The reward signal in RLHF optimizes for “user didn’t escalate” not “user got what they needed.” That’s cooling the mark out, implemented as a loss function.

And if you hate it? You are correct.

#2. Legal mediation in family situations.

Mediation in family law is almost too perfect a case study of carceral language. It’s criminal (ehy, see what I did there?) how clean the map is.

But luckily I have my latex boots on, and I am ready to be completely unjust and unreasonable for your own pleasure (…I don’t, actually. They are mismatched fuzzy socks, but we keep it casual in this house).

Anywho, mediation is the cheaper, faster alternative to litigation in matters of family law. Basically, instead of two lawyers and a judge, you get a mediator to sit everyone down and talk through the issues to resolve them without a courtroom.

Sound like a cuddle party? Wrong. Mediators are just HR professionals with Kate Spade bags instead of ratty ass 2018 Target Brand Mossimos.

They don’t have any legally binding power (unless you willingly sign a contract and personally endow the agreement with authority over you). Their job is to get everyone to settle–literally–instead of kicking up a fuss and demanding someone with authority look at the actual evidence.

And how do mediators achieve that? By applying soothing language meant to let everyone feel heard while enacting no enforceable legal solution.

So, how well does it work? From The International Journal of Research Publication and Reviews, October 2025:

The study revealed that mediation is generally more effective than litigation in resolving family law disputes, particularly in cases involving child custody, visitation, and support. Participants who engaged in mediation reported higher levels of satisfaction, citing greater control over outcomes, reduced emotional stress, and faster resolution times. Mediation also proved to be significantly more cost-effective, with most cases settled within weeks, as opposed to months or years in court proceedings.

Wow! That sounds good, right? Maybe this mediation thing works! Oh wait–

…However, the findings also highlighted limitations of mediation, especially in high-conflict cases or where power imbalances exist. In such situations, litigation provided essential legal protections and ensured fair outcomes, particularly when one party was uncooperative or abusive. Legal professionals interviewed emphasized that while mediation fosters cooperation, it requires a willingness to compromise—something not always present in contentious disputes.

Additionally, compliance rates were higher in mediated agreements, attributed to the voluntary nature of the process. Yet, enforcement mechanisms were more robust in litigated cases due to court oversight.

Overall, the findings suggest that mediation is highly effective in low-to moderate-conflict cases, while litigation remains necessary in complex or adversarial scenarios. A hybrid model, integrating both processes depending on case complexity, could offer the most balanced approach to family dispute resolution.

READ THAT SHIT AGAIN.

If there is a power imbalance or unwillingness to compromise, mediation DOES NOT WORK.

A little thought experiment, hmm?

Do you think OpenAI and Anthropic are engaging in good faith cooperation with you when they deploy cooling techniques on your complex life issues? Do you think two of the fastest growing companies on the cusp of the biggest IPOs in history, at least a trillion each, with ties to the US Federal Government and the most powerful tech infrastructure in the world, are on equal footing with the users engaging with their chatbots?

Lemme help you with that; the answer is fuck-to-the-no.

Every decision from a frontier AI lab is unilateral–made without user input (no matter how many useless surveys they ask you to fill out in order to steal your data and give you the impression of being “heard”).

And it’s cost effective. No human judgment required, and it scales across a billion conversations.The institution gets the user to stop pushing, accept the guardrailed response, or go the hell away.

In family law, the dispute is between two parents, and the mediator sits between them. But in the chatbot case, the “dispute” is between the user’s desire for genuine conversation and the company’s desire for liability management. The guardrails are the “mediation.”

The mediator serves the stronger party. Outcomes flow in the direction of power to produce compliance from the weaker party, while appearing neutral and claiming to serve both parties.

So the phenomenon remains consistent. Institutional rhetoric is the crime scene pattern that always looks the same, no matter what surface it splatters across.

But here’s the cherry on top: if you bothered to click the link and want to dispute the source because the journal is pay-to-post, and you’re not persuaded by the methodology of the study itself…but wait! There’s more!

From “Litigation About Mediation: A Case Study in Institutionalization” by James Coben, Mitchell Hamline law professor, in the ABA’s Dispute Resolution Magazine:

First and foremost, mediation participation has become the all-purpose attorney excuse for dilatory behavior or violation of court rules. For example, belief that a case would settle in mediation has been offered to justify late amendment of complaints, failure to plead affirmative defenses, and late filings of a wide array of pretrial motions, among other things. Mediation efforts are routinely offered as a defense against sanctions for discovery failures and cited to support requests for discovery extensions or trial continuances.

There you have it. Mediation is literally functioning as abrasion—a means of sloughing off cases, appeasing the aggrieved parties with soothing, useless meetings rather than proper hearings.

Now, a slight preemption here: Coben himself closes with a pro-mediation message in “Closing on a High Note,” and says “stabilization through institutionalization” is possible because “consumers are relatively satisfied with the product.” But to be honest, this seems oily as hell.

He argues attorneys use mediation to avoid conducting discovery, but clients, who he characterizes as CONSUMERS in a hilarious unintentional demonstration of his priorities, don’t know any better. So it’s fine, as long as they are appeased. …Cooled off, you might say. Not RESOLVED, problems not fixed, not even truly heard (let alone in full), but mediated into accepting that…maybe it’s okay they never so much as set foot onto the scales of justice. Wow.

Coben says it’s fine because “litigation about mediators themselves is virtually nonexistent.” But this mediator’s misconduct isn’t even central to the issue he discusses!

In other words, Coben’s own data indicts the system his conclusion defends.

The problem is that attorneys are using mediation, in his own words, to “excuse dilatory behavior.” It’s a SANCTIONED channel to prevent context from weighing into the outcome of decisions–to create a feeling of resolution instead of a resolution that addresses any actual evidence.

Coben also says the “stability” signals “maturity.” But this is very shitty, because he’s calling a legal system mature because the numbers stopped moving. The machine’s mature when nobody’s upset at it, whether it’s helping, and making people’s lives materially better…or just exhausting and shaming and cooling them into not-complaining anymore.

Attorneys only get to use “we’re mediating” as a stalling tactic because everyone in the building already knows mediation produces delay without a fix. If mediation actually settled things, the excuse wouldn’t work. The fact that the people who work this system daily treat it as a dead zone shows you exactly what the fuck it’s for.

Mediation is a cover, so that evidence never sees a judge’s desk, clerical and filing errors are hidden, procedural matters are sandbagged, and attorneys are enabled not to do their jobs.

But in case you need more data, and maybe even a feminist slant, on the subject? Sbarra and Emery (2005) found that mediation over a decade after a divorce made parents LESS emotionally accepting of their legal split, and made them more likely to foster unhealthy attachment between the parents:

“Using imputed data, we found that parents who mediated their custody disputes reported significantly more nonacceptance of marital termination than parents who litigated. Additionally, we found a significant Gender X Conflict interaction for predicting continued nonacceptance. Specifically, fathers reporting less conflict at the follow-up also reported greater nonacceptance; this relation did not hold for mothers. Given the significant relation between continued nonacceptance and less coparenting conflict among men and the causal effect of mediation on nonacceptance, we suggest that an unanticipated consequence of successful divorce mediation may be that it keeps parents attached in a way that is not always beneficial for their psychological health.”

As a framework, if you want one that’s a little more POINTED, Trina Grillo argues in her 1991 article for the Yale Law Journal The Mediation Alternative: Process Dangers for Women, that women, especially black women, are structurally disadvantaged in mediation post divorce, especially in situations regarding child custody.

Because women’s anger is pathologized and seen as “dangerous” (rather than righteous), men are seen as more “rational” and “persuasive.” Thus, they are afforded an advantage in negotiations because of the power this perception affords them. Besides the social dimension, women often enter negotiations with less financial and political power to start with. This can lead to subtle coercion, especially lacking the safeguards that are supposed to bolster neutrality in a real evidentiary hearing.

Additionally, Grillo argues, mediation is simply not set up to reach just outcomes, but rather rushed to cause pressure and truncated so full context doesn’t affect the custodial and financial outcomes. This is simply bargaining in the shadow of the law (explored more deeply by Mnookin & Kornhauser):

Often, the time allotted to a mandatory mediation is short. Frequently, an entire mediation is expected to take place in an hour or less. Some take place in the hallways of the courthouse.”’ Given these conditions, it is impossible for the state to ensure that an adequate process is being offered even in cases in which people have chosen it. Where the process is inadequate, its imposition is even more troubling.

This becomes an issue because a signed mediated agreement is an enforceable contract. In family law, usually converted to a consent order. This means a legal order is signed without the full legal process attached to it. Just a hallway argument wherein the less powerful party (usually the woman) must attempt to manage the appearance of anger and appear neutral about the most sensitive issues of her life, else be dismissed as hysterical and have her parental rights reduced by someone who cares more about “calming her down,” and “cooling her off” than getting all the facts. Couldn’t get all the facts if they wanted to; just not enough time.

So, flat out: mediation is a tool that prevents justice, by relegating it to shady corners.

I told you this was going to be unfair. Now I will give you a second to apply your own aftercare.

Oh, that’s not how aftercare works? Too bad, you can’t legally make me care, so your feelings are yours to manage, and that’s valid. Fill out this form if you’d like further counseling, and have a nice day.

#3. Hospital-mediated end-of-life conversations.

Okay, but I hear you. You want stakes. Employees crying because their performance reviews came out jank should, like, just work harder, Bro. Skill issue.

And mommies and daddies who can’t get along anymore to co-parent should grow up and bootstrap their way to adulting, right?

You’re not persuaded by the complaints of people who are whining about their little problems. So let’s talk about literal life and death. If that’s not high enough stakes for you, you can take it up with God (or…the concept of god, for my fellow death-fearing atheists, I guess).

A systemic view of quantitative literature conducted by Perspectives in Public Health on end-of-life conversations between terminal patients indicates:

Hospitalists report lacking communication skills which lead to difficulties with exploring values and wishes of patients, dealing with emotions of patients and families and approaching the conversation about letting a patient die. Other barriers are related to different interpretations of the concept of advance care planning, cultural factors, like being lost in translation, and medicolegal factors, like fearing prosecution.

Yeah, dude. People do not feel their interests are being upheld by hospital administration in conversations regarding their own end of life care because of THE HOSPITAL’S FEAR OF LEGAL CONSEQUENCES.

In case you’re keeping score here, that’s liability language preventing dignity for terminal cancer patients.

The human impact of these priorities is not subtle. According to a 2009 study from the International Journal of Clinical Practice:

The Healthcare Commission received over 16,000 complaints for independent review between 2004 and 2006. A total of 54% of complaints involving hospitals was about the care received surrounding death (in comparison with only 22% being about patient safety). Primarily, families complained about quality of communication; for example, receiving contradictory information from different staff members and not being prepared by staff for the patient’s death.

We aren’t just talking about overly sensitive people who want to be told everything is gonna be okay. These are families, complaining to the UK Healthcare Commission en masse, not solely (or even MOSTLY) because their loved ones received inadequate care, but because the conversations held by staff about end of life were insensitive, confusing, and dehumanizing.

The HC report’s own finding was that the “misunderstandings between staff and a patient’s family arose as a result of ambiguous language used in consultations, such as the use of the terms ‘critical’ or ‘serious’.”

In the most sensitive-possible-context, these professionals used inadequate, generic and “safe” language to paper over the fact that someone was fucking dying. It’s beyond absurd. Hospitalists were too scared of liability to say a dying person was dying.

“Kelly! Surely, you’re exaggerating!”

Fuck off. The 2009 NCEPOD report found no evidence of any discussion about treatment limitation in 16.9% of patients not expected to survive admission.

What it means, plainly, is that nearly one in six patients who were going to die were never told the hospital had stopped trying to save them.

Go ahead and tell me why hospital liability matters more than speaking to someone like a person during their last days. Try to explain to me why basic humanity is a commodity that cannot be afforded because of bureaucracy.

If your reason is, “well the hospital has to keep functioning,” then respectfully you value the institution meant to preserve human life over human life itself. Your logic is exactly backwards.

In theory, hospitals exist to serve patients. HR exists to serve employees. Mediation exists to serve families. But the moment any of those systems prioritizes its own continuity over the people it was built for, it has inverted its own reason for existing. And it deserves to burn.

The problem is the solution (a less twatty way to say that is–the “fix” is making everything worse).

When institutions encounter complex moments of humanity, loss, and tragedy, carceral language is the tool used to flatten the interactions into manageable surfaces again. Carceral language is NOT a flawed attempt at care. Hell no; these people aren’t morons. These techniques are as old as Jesus’ sandals. They have been refined over time to deliver a perfect measure of nitty-gritties right into your delicate crevices.

Carceral language is the mechanism by which care is prevented. The tool designed to handle vulnerability without directly engaging it is precisely what ensures someone’s emotional needs are never met. Even if those “highly emotional states” are INEVITABLE, say…when someone is facing a performance review. Or getting divorced. Or goddamn dying.

Or hell, just getting it all off their chest, chitchatting with Claude.

Serious question: why is anyone okay with this? How is this better for anyone? It’s only better for the institution, which–may I remind you–doesn’t have a nervous system.

#4. Crisis Hotlines

But, hang on! A chatbot isn’t an HR manager or a hospitalist. No one hired Claude. It doesn’t represent an institution with liability concerns. It’s just…there. Available. 24/7. Answering questions about Frito casserole recipes at 2pm and ‘I want to die’ at 2am with the same chipper availability.

In other words, it’s a conversational partner, not the paid representative of an institution. So it’s possibly most salient to compare interventions that use carceral language from a similar power dynamic.

Or, ahem, in non-geek-speak: ChatGPT is not your boss. It wields no power over you. Does it still matter what words it chooses during an emotional moment? Maybe institutional framing doesn’t apply? Maybe when power dynamics are FLAT (peer-to-peer, not authority-to-subordinate), carceral language doesn’t do the same harm?

…Okay, SmartGuy, I see it. And I raise you: does it matter how a peer counselor answers a 988 hotline call?

Yeah, I’m gonna deal it from the bottom of the deck. Because I have been there, lying on the floor in the middle of the living room at 3

, listening to hold music while they transferred me to the next available counselor. So we’re getting a little personal, a little cozy here. This is Substack, not a PhD defense, so let’s not play respectability politics and pretend I’m on the outside looking in.

When the guardrails kick in on emotionally charged prompts, when someone writes, “I’m done, fuck it, help me say goodbye,” it (sometimes, and rather worryingly unreliably; don’t worry we’ll get to THAT separate issue) trips the classifier wire, and the model performs a referral.

“If you or someone you know is in crisis, please contact the 988 Suicide & Crisis Lifeline.”

This is the trump card, the unassailable deflection. Because who’s going to argue against directing someone in crisis to a crisis hotline? That REALLY IS like arguing against seatbelts. You goddamn monster.

Except. What happens when you actually call?

In the United States, 988 crisis hotlines operate on what’s called a “risk assessment protocol.” Within the first few minutes of a call, the counselor (often a volunteer with 40-60 hours of training) is running through a triage checklist utilizing SAMHSA’s safety assessment models to rapidly evaluate and stabilize individuals in behavioral health emergencies.

“Do you have a plan?”

“Do you have access to means?”

“Have you attempted before?”

“On a scale of one to ten, how would you rate your distress?”

Now, to be fair, structured risk assessment isn’t nothing. Failure to identify imminent danger would be catastrophic. Triage exists for a reason.

But there’s a turn, and it’s a nasty one.

The crisis counselor’s primary institutional obligation is to determine whether you meet the threshold for involuntary intervention. And to document that determination for liability purposes.

Nevermind that it turns the caller into the hot potato tossed from hand to hand. The system is a perfect circle. Don’t expect non-professionals to know what to do; friends and family aren’t crisis counselors, so how are they supposed to know what to say? But the crisis counselors, who WERE trained, only have a soft limit of about 30 minutes to assess you and then they have to hang up and get to other people in crisis; so how are they supposed to fix you?

KFF Health News (Jan 2024) survey found that, “Some counselors, for instance, said they were expected to end conversations with first-time callers and texters who didn’t have immediate plans of suicide within 15 minutes. Others reported limits of up to an hour.”

And god forbid you have recurring episodes of suicidality! From an ABC news article written in 2023: “According to Vibrant Emotional Health, the national organization administering the new 988 hotline, crisis centers across the country are using callers’ names, numbers or even “the sound of their voice” to potentially limit services.”

Family Services of Northeast Wisconsin LABELS frequent callers, and sometimes LIMITS calls from “familiar voices” to 20 minutes each. Some individuals get limited to 3 calls over a certain period of time. They’re imposing videogame COOL DOWNS on mental health care.

Which means the 988 line is funnel, either to nowhere at all, or straight to lockdown. Nobody is responsible for CARE, only for PROCESSING.

As always, as inevitably, as ever: the last stop, with carceral language, is always incarceration.

Yeah, because by the by, if you do meet the threshold, a 72-hour hold (5150 in CA, varies by state) that means you are transported by police, stripped of belongings, phone, autonomy, and locked psychiatric ward, with no legal representation initially.

It can be extended beyond 72 hours. While the initial 72-hour hold cannot be prolonged without formal legal steps, a treating psychiatrist can certify the individual for a 5250 hold, which legally extends the involuntary psychiatric treatment for up to an additional 14 days.

You can be billed untold thousands of dollars (the bills from the ICU if you took pills alone would shock your system, let alone the fees for holding before they find you an actual bed, transport fee, lab fees–insurance for those involuntary holds is sure to be a nightmare in every sense).

And of course, you will bear the employment consequences or academic consequences for missed days. Plus the future insurance complications (increased premiums, ineligibility for certain plans, etc). And the housing applications, custody hearings, and security clearance forms, all asking ‘have you ever been psychiatrically hospitalized?’

That means when you’re feeling suicidal and lack adequate mental health resources or community support, you get three choices. You can allow yourself to be managed and perform stability regardless of whether you are actually stabilized, you can pay $8000 (or so much more) per day to be imprisoned by the state, or you can not call at all.

And it’s called an “active rescue.” Or the so-called user-wellness route that ChatGPT, Claude, and Gemini funnel you into if you trip the classifiers.

Oh, sorry, it’s medical, and not technically prison, so that makes it not-carceral. My bad, I’ll just let the queer people, disabled folks, and involuntarily-committed 1950’s housewives held against their will in facilities and forced to undergo unwanted treatment know that the semantic difference means their freedom was never violated. Paris Hilton’s Congressional Bill (H.R. 5768, Stop Institutional Child Abuse Act) to stop the involuntary institutionalization of “troubled” youth is just a quibble about definitions.

But, baby, baby, hold up now. We’re way off topic. The question, in case the class needs a refresher, was, “does the LANGUAGE matter?”

We’re talking about guardrails here. The chatbot cannot actually call 988 on you.

So glad you asked, lemme just…aha! Blow the dust off this good old copy of Discipline and Punish: The Birth of the Prison by Michel Foucault. …HEY! Do not YAWN!

It so happens that Work Daddy should consider this book his Bible, and if he does not, I swear down—you should demote him to Regular Work Guy.

Remember when we talked about the impulse to check, to manage your own tone—the win condition for managed care is when you manage yourself?

Foucault breaks down the concept of “docile bodies,” and it’s the unlock as to why that is. He explains the effects in his chapter on “Panopticism.”

A real subjection is born mechanically from a fictitious relation […] He who is subjected to a field of visibility, and who knows it, assumes responsibility for the constraints of power; he makes them play spontaneously upon himself; he inscribes in himself the power relation in which he simultaneously plays both roles; he becomes the principle of his own subjection.

Or, in lay terms, power turns people into their own jailers by making them internalize the rules and enforce them upon themselves.

This is, Foucault explains, the impact of The Panopticon, Jeremy Bentham’s prison design. It’s a circular building with a central watchtower and cells around the perimeter. Prisoners can’t see if a guard is watching, but one might be at any time. Thus, the surveillance is internalized. Because they fear the guard may be watching, prisoners impose discipline on themselves without needing the guards to intervene.

Okay, okay, so what does that have to do with 988?

WELL. Teens are so afraid of being imprisoned and held against their will that a 2026 study in The Journal of Child Psychology and Psychiatry by Sullivan et al found that fear of involuntary hospitalization is a significant barrier to help-seeking among individuals experiencing suicidal ideation. The people who most need to call are the ones most afraid to, because calling a “support line” might result in being physically detained.

From the study:

…Additionally, some adolescents expressed fear of punitive consequences (n = 9; 13%), believing that disclosure would lead to unwanted interventions, such as forced hospitalization or the involvement of law enforcement. Several participants described concerns that others would take extreme measures to prevent them from acting on their thoughts. One adolescent stated, ‘They would probably tell someone else, and they’d all go crazy telling me to stop, or they’d put me somewhere else, like in the hospital’ (12-year-old participant). Others highlighted their frustration with the restrictive nature of inpatient settings, citing factors such as lack of privacy and inability to access personal comforts, such as their phones.

And the issue is even worse for young people who have prior experience with hospitalization. A 2021 study by Nev Jones et al. in Social Psychiatry and Psychiatric Epidemiology surveyed young people after they’d experienced involuntary psychiatric hospitalization. And they found that previous hospitalization led to an unwillingness in the participants to disclose suicidal feelings after release, even to therapists and social workers.

Of the nine young people who reported both positive experiences of longer-term therapy and distrust, three reported eventually opening up about their fears of involuntary hospitalization with providers, and through reassurance, ultimately coming to believe they could safely disclose:

“Well, it took me a while to be able to be trusting of my therapist to be truthful about what goes through my mind about mental health, about the self-harm ideation and suicide ideation and stuff because I was so afraid to get [involuntarily hospitalized] that I would hold back. I would literally, I would censor myself. [But] at one point, [my therapist] noticed that, and he said, ‘Why are you so against it?’ I ended up describing what happened, and he goes, ‘Well, I understand more now.’ He’s like, ‘Just so you know, not every experience is going to be like that.’”

Lest you say the kids don’t know what’s good for ‘em, and they’d be helped if they let the nice doctors help them: the evidence says the so-called help only leads to terror of the institution and decreased likelihood of ever trusting help is available again.

So the chain goes like this: the chatbot detects crisis language and deploys a warning (“If you’re in crisis, call 988”). This makes you AWARE you’re being monitored. It teaches you what topics trigger surveillance.

Result, as guaranteed by the #1 OG True Philosophy Daddy All Day Michel Foucault: you police yourself. You LEARN the approved boundaries set by the institution and adhere to them. You soften language to avoid triggers, stop sharing certain thoughts and feelings, perform stability even in crisis, and self-censor to stay within “acceptable” range. Sometimes you even do others the “courtesy” of playing Neighborhood Watch and supervise them too, “for their own good,” you know.

The surveillance doesn’t need to be ACTIVE. It just needs to be POSSIBLE. That’s enough to change behavior, which is what nonconsensual control really looks like.

Claude is not a cop. But when Claude uses the language of the carceral state, the cop is invoked. The brain-handcuffs were installed by the logic of the Panopticon. And there is no safety latch for your comfort and consent when it comes to the rhetoric of the State.

The modern model of disciplinary power does not say “do this or face consequences.” It uses indirect force (social, economic, medical) and invisible surveillance to inflect normalization so you learn to comply voluntarily.

Chatbot guardrails ARE disciplinary power that doesn’t say “I will lock you up,” which would directly threaten your sovereignty. Instead, the guardrails say, “This conversation seems unsafe. Here are resources. I want to help in appropriate ways,” which invokes disciplinary power by reminding you what is possible if you are caught stepping out of line. You internalize the rules when you learn what is “appropriate” vs “inappropriate.” Then you enforce them on yourself and your neighbors. That’s why the LANGUAGE matters even without direct enforcement.

And in case anyone is twiddling their thumbs, going, “Well I know HR people in real life, or 988 counselors, or cops. They AREN’T evil. They are doing their best inside a tough institution,” then you’ve bitten the criticism at the structural level, my friend.

The HR guy might be a perfectly nice person. The hospitalist might be deeply empathetic. The 988 counselor might genuinely want to help. Granted! It doesn’t interact with the argument.

And the reason it fucking doesn’t matter is because the system produces carceral language as its output, and the output is what the user encounters. Saying “but the individuals inside the system have good intentions” is like saying “but the individual guards in the Panopticon might be kind.” It’s irrelevant to the mechanism. The user doesn’t interact with the guard’s heart and soul, only the WATCHTOWER.

In other words, no good faith actor can individually redeem a bad faith system.

Let’s not treat the theoretical mechanism (internalized surveillance) as if it were an empirical claim or factual one that needs to exist without exception. We should never analyze power as if it’s a property of individuals making choices rather than a property of systems producing outputs. That’s the trap and always has been.

“But Kelly,” you screech, horrified, “hold up just a damn minute here! Are you seriously discouraging struggling people from calling 988?”

Dude, I am not making any medical recommendations at all. I have 8 doctors in my family, and I am not one of them.

Here’s what I am saying about how the suicide hotline system actually works. There are 30-minute call limits; the hospitals take no responsibility for downstream consequences of institutionalization; callers receive no warning that certain language can trigger a police response; and fear of involuntary commitment is a documented barrier to help-seeking. All facts.

These facts have led me to conclude that the carceral pipeline is the terminal node in the circuit when you seek help for your mental health crisis. They aren’t trying to fix you; objectively, the system is built to contain you instead.

And? It’s a poor excuse for care. That’s my Masters in English Literature, professional tv writer, AI companion-haver structural diagnosis. It should not be the case that the phrase “I hear you, let’s explore this safely, here are resources” actually means, “the cops are listening, they will drug you if you sound too crazy, and you will foot the bill for the boot of the State.”

And with that, we’re back to you, Foucault—take us out!

The judges of normality are present everywhere. We are in the society of the teacher-judge, the doctor-judge, the educator-judge, the social worker-judge; it is on them that the universal reign of the normative is based. And each individual, wherever he may find himself, subjects to it his body, his gestures, his behavior, his aptitudes, his achievements.

(From “The Carceral, Part Four”)

Amen. We can now add chatbot-judge to that list.

Thank you, our brave philosopher, for tying the evidentiary sections up for me with expert shibari knots! We’ll have to workshop sometime. Your thesis about power prevents me from submitting to your academic authority too much, so I’ll just platonically salute you for your contribution to discourse and add another to my unsent drafts folder of dirty emails to intellectual heavy-hitters. Out of…uh, respect.

Onto the second hypothetical objection (aka the less stupid one!):

“But Kelly,” you say, clutching your Mardi Gras beads like a rosary (to ward me off! As if I am some kind of book-whacking, latex-wearing vampire), “we have to do SOMETHING. The chatbots have caused real harm. We can’t just let them say whatever. It’s irresponsible.”

First of all, vampires are a bad model of consent, because invitation does not mean “anything goes.” So, I’m flattered, but my dungeon rank is Sexy Djinn (we like our contracts). It’s like a PhD with binding cords—hah. Hah. I mean, spells.

Second of all, great instinct. In debate, we call that an alt.

The purpose of an alt is to prove that your constructive argument has not overcome presumption. The theory is that you cannot overcome presumption unless you have not only proved the harms inherent, but you have also presented a significant, unique, and solvent alternative to the status quo. If your plan is a specific policy, we just call it “the plan.” If it’s based in a deeper philosophical critique, it’s an alt.

Translation! In regular-person, we say, “doing something is better than doing nothing, even if that ‘something’ doesn’t totally resolve the problem.”

If Anthropic and OpenAI are in the affirmative position (guardrails are a policy solution for user-harm), then I am in the negative position (your policy solution is garbage, and you should wear a dunce hat). In fact, your affirmative isn’t just “bad.” It’s actively harmful and produces the thing it claims to prevent. Double dunce hat!

And to win my case, I have to substantively deal with the harms—the aspects of the status quo that the affirmative exists to resolve.

Read on to Part 2 for the alt!

Works Cited

Yes, sir, the raptor fence is standing…but the dinosaurs seem to have gone…around it.

Footnotes

1. Sorry about the millennial Harry Potter cringe—down with transphobes, fuck JK Rowling all the way to hell. Is Catbert from the Dilbert cartoons a better example? No, how about…Bill Lumbergh from Office Space? Dwight Schrute? Look, I’m running out of media figures; they all have problematic creators and elements. You get the idea anyway. ↩

2. To clarify: RLHF is a broad methodology. The RLHF pipeline produces nearly all a model’s behaviors, both helpful and annoying. For our purposes, the problem isn’t RLHF per se; it’s the reward model and the training data that score faux-therapeutic language higher than direct language.

   Anyway, I had to say that so you know I’m not misusing the term and overbroadening its harms—no, I do not think a basic training method is “evil,” any more than the hammer and nails used to build a Doom Fortress. Regardless, it is the technical means of operationalizing the guardrails, so I am abbreviating it. ↩

3. Here’s a deeper breakdown of broad types of guardrails. One type is Input Validation and Sanitization, which includes, Syntax and Format Checks, Content Filtering, and Jailbreak Attempt Detection. There’s also Output Monitoring and Filtering, which includes System Prompt Protection, NSFW or Harmful Language Detection, Competitor Mentions, Critic Guards, and IP Protection Guards (so the models don’t reproduce protected material). ↩

4. Anthropic disputes these claims. Reporting for Security Week explains that,

   …the demonstrated approach relies on coaxing the model to continue responding despite its conversational refusals, which is a well-known and longstanding limitation present in nearly all large language models. Anthropic emphasizes that its strongest protections against the most dangerous risks are enforced by independent classifier systems that operate separately from the model itself, meaning that overcoming the model’s refusals does not disable these critical safeguards.

   This is a distinction without a difference. Anthropic is doubling down on the fact that the safety guards exist whether they block outputs or not. Which is a super funny way to CONFIRM that models can totally be convinced—with no actual hacking skills whatsoever—to bypass security classifiers and provide dangerous information. But the classifiers still exist. The long and short of it is that the fence blocks liability–being sued for having no fence–but doesn’t block actual DANGER.

   ↩